Microsoft Azure AI Fundamentals AI-900 (AI-900) — Questions 676–750

The requirement focuses on protecting patient identities by not storing or transmitting PII, which directly addresses privacy and security, not fairness.

The requirement focuses on protecting patient identities by not storing or transmitting PII, which directly relates to privacy and security, not inclusiveness. Inclusiveness is about ensuring the system works for all user groups, not about data protection.

The requirement focuses on protecting patient identities by not storing or transmitting PII, which directly relates to privacy and security, not reliability and safety. Reliability and safety concern system accuracy and harm prevention, not data protection.

Fairness would be correct if the question described an AI system that must avoid bias against certain patient groups, such as ensuring equal treatment recommendations across demographics.

A question asks: 'An AI system for healthcare must provide equal quality of service across different demographic groups. Which principle is most directly addressed?' In that scenario, inclusiveness would be correct.

A question asking which principle ensures an AI system for medical diagnosis consistently produces accurate results and avoids causing patient harm due to errors or failures.

Candidates may confuse privacy with fairness because both involve ethical handling of data, but fairness is about equitable outcomes, not data protection.

Candidates may confuse inclusiveness with privacy because both involve ethical considerations in AI, but inclusiveness is about fairness across groups, not data protection.

Candidates may confuse 'safety' with data security, or think that protecting patient data is part of ensuring system safety, but safety here refers to operational reliability, not data confidentiality.

Underfitting would cause poor performance on both training and test sets, but here training accuracy is high (98%) while test accuracy is low (72%), indicating the model memorized training data rather than failing to learn patterns.

Data leakage typically causes overly optimistic performance on both training and test sets, not a large gap between them. Here, the high training accuracy and much lower test accuracy indicate overfitting, not leakage.

Class imbalance typically causes poor performance on minority classes, not a large gap between training and test accuracy. The 98% vs 72% discrepancy indicates overfitting, not imbalance.

A model achieves low accuracy (e.g., 60%) on both training and test sets, suggesting it is too simple to capture underlying patterns in the data.

A model achieves 99% accuracy on both training and test sets, but fails in production. The question would describe that future data contains a feature (e.g., 'patient ID') that was inadvertently used during training, causing unrealistic performance.

A question where a model shows high accuracy on a test set but performs poorly on a specific minority class (e.g., fraud detection with 99% overall accuracy but 0% recall on fraud cases) would make class imbalance the correct answer.

Candidates may confuse 'low test accuracy' with underfitting, not realizing that high training accuracy rules out underfitting.

Candidates may confuse the symptoms of overfitting (high training, low test accuracy) with data leakage, as both can lead to misleadingly high training accuracy. They might think the test accuracy drop is due to leakage rather than memorization.

Candidates may confuse any performance issue with class imbalance, especially if they recall that imbalance can cause misleading accuracy, but here the key symptom is the train-test gap, not skewed class performance.

Precision measures the proportion of positive predictions that are correct, but in this case the model predicts 'cat' for all images, so precision for 'cat' is 95% (since 95% of images are cats), which is misleadingly high and does not reveal the model's failure to detect dogs.

Recall measures the proportion of actual positives correctly identified, but here the model predicts all images as 'cat' (the majority class), so recall for 'dog' (the minority class) is 0%, which is poor. However, recall alone does not account for precision, and in this imbalanced scenario, the F1-score (harmonic mean of precision and recall) provides a more balanced evaluation.

Accuracy is misleading on imbalanced datasets because a model can achieve high accuracy by always predicting the majority class, as in this case where the model predicts 'cat' for all images and achieves 98% accuracy despite having zero predictive power for dogs.

When the cost of false positives is high, such as in spam detection where marking a legitimate email as spam is unacceptable, precision is the key metric to minimize false positives.

In a scenario where the cost of missing positive cases is high (e.g., detecting rare diseases), recall is prioritized. For example, if a model must identify all instances of a rare cancer, even at the expense of false positives, recall would be the key metric.

Accuracy is the correct metric when the dataset is balanced (e.g., equal number of cats and dogs) and the cost of false positives and false negatives is equal. For example, in a question where the test set has 50% cats and 50% dogs, accuracy would be appropriate.

Candidates may think precision is important because it focuses on correct positive predictions, but they overlook that in this imbalanced dataset, the model's high precision for the majority class masks its poor performance on the minority class.

Candidates may think recall is important because the model fails to detect dogs (the minority class), but they overlook that recall for the majority class (cats) is perfect, and without considering precision, recall alone doesn't reveal the model's trivial prediction strategy.

Candidates often default to accuracy because it is the most intuitive and commonly used metric, and they may overlook class imbalance or fail to consider that high accuracy can be achieved by a trivial model.

Image classification assigns a single label to the entire image, but the question requires detecting the presence and location of multiple products on shelves, which object detection provides.

OCR extracts text from images, but the question requires detecting the presence and location of any product (e.g., a box, a bottle), not reading text on labels.

Semantic segmentation assigns a class label to every pixel, which provides pixel-level masks but is overkill for simply detecting the presence and location of products; object detection with bounding boxes is sufficient and more efficient for this use case.

A company wants to classify shelf images as 'stocked' or 'empty' without needing to locate individual products. Image classification would label the whole image accordingly.

A company needs to read expiration dates or batch numbers from product labels on shelves to verify freshness or traceability, without needing to locate the products themselves.

A medical imaging company needs to precisely outline tumor boundaries in MRI scans to calculate tumor volume. They require pixel-level classification of each region, not just bounding boxes. Semantic segmentation would be the correct choice.

Candidates may confuse 'detecting presence' with classifying the image content, not realizing that location information is also required.

Candidates may think product labels contain text that OCR can read, and mistakenly believe that detecting text on products is equivalent to detecting the products.

Candidates may confuse semantic segmentation with object detection, thinking that 'detecting presence and location' requires pixel-level precision, not realizing that bounding boxes are adequate for location without fine-grained outlines.

Increasing the temperature to 1.0 makes the model more creative and random, which increases the likelihood of generating fabricated or incorrect information, contrary to the goal of minimizing fabrications.

Fine-tuning on the knowledge base documents does not inherently prevent the model from generating plausible-sounding but incorrect information; it can still hallucinate or produce fabrications, especially if the training data is incomplete or the model overgeneralizes.

A system message telling the model to never make up facts is a form of prompt engineering, but it does not provide the model with actual correct information from the knowledge base. Without relevant context, the model may still hallucinate because it lacks authoritative source material to ground its responses.

In a scenario where the goal is to generate diverse, creative responses (e.g., a story generation tool) and factual accuracy is not required, increasing temperature to 1.0 would be appropriate to encourage varied outputs.

When the goal is to adapt a base model to a specific domain or style (e.g., legal or medical terminology) using a fixed dataset, and the primary concern is improving accuracy on that domain rather than preventing hallucinations from external documents.

This option would be correct in a scenario where the model already has sufficient accurate knowledge (e.g., from pre-training or fine-tuning) and the goal is to enforce strict adherence to factual responses without relying on external data retrieval. For example, a question about reducing off-topic or speculative answers in a general-purpose chatbot where the model's training data is already reliable.

Candidates may think that higher temperature forces the model to rely less on its training data and more on provided context, but in reality it increases randomness and hallucination risk.

Candidates may believe that fine-tuning the model on the knowledge base will 'teach' it the correct information, thereby eliminating fabrications, but they overlook that fine-tuning does not guarantee grounded responses and can still lead to hallucinations.

Candidates may think that a strong instruction in the system message is enough to prevent hallucinations, underestimating the model's tendency to generate plausible-sounding but incorrect information when it lacks specific factual grounding.

Image classification assigns a single label to an entire image, but the task requires detecting and counting multiple people in a video feed, which is a spatial detection problem.

Object detection identifies and locates objects in an image but does not specifically count people; it would require additional logic to filter for 'person' class and track across frames, whereas People Detection is optimized for counting people in crowds.

OCR is designed to extract text from images, not to detect or count people. The question requires counting people entering a store, which is a people detection task, not text recognition.

A question asking which capability to use for categorizing a store image as 'busy' or 'empty' based on overall scene content would make image classification correct.

A question asking: 'Which Computer Vision capability should be used to identify and locate multiple product types on a shelf in a retail store?' would make object detection correct, as it can detect various object classes with bounding boxes.

A question asking which Azure Computer Vision capability should be used to read text from store signage or receipts from live video feeds would make OCR the correct answer.

Candidates may confuse image classification with object detection, thinking that classifying an image as 'person' is sufficient for counting, but classification does not provide locations or counts.

Candidates may think object detection can count people by detecting 'person' objects, overlooking that Azure's People Detection is a specialized prebuilt model for accurate people counting in crowded scenes.

Candidates may confuse OCR with general image analysis capabilities, thinking it can 'read' or identify objects in a scene, or they may not be aware of the dedicated People Detection feature.

Temperature adjustment controls randomness of output, not style consistency. It cannot learn from 50 sample descriptions to match a specific brand voice.

Prompt engineering with detailed instructions can guide the model's output but cannot specialize the model to consistently match a specific brand voice across many descriptions; it lacks the ability to learn from the provided 50 samples to internalize the style.

Content filtering is used to detect and filter harmful or policy-violating content, not to specialize a model's output style or tone.

When the question asks how to control creativity vs. determinism in generated text, e.g., 'Which parameter should be lowered to make model outputs more focused and repetitive?'

When the question asks for a method to guide the model's output without additional training, such as generating text in a specific tone for a one-time task where the user can provide explicit instructions in the prompt.

A question asking how to prevent an Azure OpenAI model from generating offensive or inappropriate content in a customer-facing chatbot would have content filtering as the correct answer.

Candidates may confuse temperature with fine-tuning, thinking that adjusting a parameter can adapt the model to a specific style without additional training.

Candidates may overestimate the power of prompt engineering, thinking that detailed instructions can achieve the same level of style consistency as fine-tuning, especially when they are familiar with prompt-based techniques but less experienced with fine-tuning.

Candidates may confuse content filtering with model customization, thinking it can shape output style by blocking unwanted patterns.

Using a smaller test set does not address overfitting; it reduces the reliability of error estimates and may hide overfitting, but does not penalize large coefficients.

Feature scaling alone does not penalize large coefficients; it only normalizes feature ranges. Overfitting caused by large coefficients requires regularization, not scaling.

Increasing the number of training epochs does not penalize large coefficients; it can actually worsen overfitting by allowing the model to fit the training data even more closely.

If the question asked how to reduce variance in model evaluation due to limited data, using a smaller test set (or cross-validation) might be considered, but typically a larger test set is preferred for stable estimates.

When training a model like SVM or k-NN that is sensitive to feature magnitudes, and the question asks for a preprocessing step to ensure all features contribute equally to distance calculations.

In a scenario where a model is underfitting (high error on both training and test data) and training is stopped too early, increasing epochs can help the model converge to a better solution.

Candidates may think that a smaller test set will make the test error appear lower, confusing evaluation strategy with regularization.

Candidates may confuse feature scaling with regularization, thinking that scaling coefficients indirectly reduces their magnitude, or they may know scaling is important but misapply it to overfitting.

Candidates may think that more training always improves performance, confusing the concept of convergence with regularization, and not realizing that overfitting is exacerbated by excessive training.

The model already overfits with 500 samples; adding more data would help generalization but is not the most direct technique to address overfitting caused by high dimensionality relative to sample size.

Increasing model complexity by adding more layers would exacerbate overfitting, especially given the small sample size relative to the number of features, making the model even more prone to memorizing noise.

Training for more epochs would likely increase overfitting further, as the model already memorizes the training data (99% accuracy vs 68% test accuracy). More epochs exacerbate overfitting rather than reduce it.

When a model underfits due to insufficient data (e.g., low accuracy on both training and test sets), increasing training data can improve performance by providing more representative examples.

If the model were underfitting (e.g., low accuracy on both training and test sets) and the dataset were large enough to support a more complex model, adding layers could help capture more complex patterns.

When the model is underfitting (e.g., low training and test accuracy) and training loss is still decreasing, training for more epochs can improve performance. For example, a neural network with high bias may need more epochs to converge.

Candidates know that more data often reduces overfitting, but they overlook that the question asks for the most direct technique given the specific context of many features and few samples.

Candidates may think that more complex models always improve performance, overlooking the bias-variance tradeoff and the risk of overfitting with limited data.

Candidates may think that more training always improves accuracy, not realizing that overfitting is already present and more epochs will worsen generalization.

Sentiment Analysis determines the overall positive, negative, or neutral tone of text, but does not extract specific aspects like 'cleanliness' or 'staff friendliness' mentioned in reviews.

Entity Recognition identifies named entities like people, places, or organizations, but the question asks for extracting frequently mentioned aspects (e.g., 'cleanliness'), which are key phrases, not predefined entity types.

Language Detection identifies the language of text (e.g., English, Spanish), not the topics or aspects mentioned in guest reviews. The hotel chain needs to extract specific aspects like 'cleanliness' or 'staff friendliness', which requires Key Phrase Extraction.

A company wants to automatically classify customer feedback as positive, negative, or neutral to track satisfaction trends over time, without needing to identify specific topics.

A company wants to automatically extract names of competitors, product names, or locations from news articles using a prebuilt Azure AI Language feature without custom training. Entity Recognition would be correct here.

A multinational company receives customer feedback in multiple languages and wants to automatically route each review to the appropriate language-specific team for processing. Language Detection would be used to identify the language of each review.

Candidates may confuse sentiment analysis with aspect-based analysis, assuming that identifying sentiment also reveals the topics being discussed.

Candidates may confuse 'aspects' with 'entities', thinking that aspects like 'cleanliness' are entities, but Entity Recognition focuses on predefined categories (e.g., Person, Location) rather than open-ended key phrases.

Candidates may confuse 'detecting aspects' with 'detecting language', or think that language detection is a prerequisite for further analysis, but the question explicitly asks for identifying aspects, not languages.

Temperature controls randomness of token selection, not topic repetition. Increasing temperature makes outputs more random but does not specifically penalize repeated topics.

Top_p (nucleus sampling) controls the diversity of word choices by limiting the cumulative probability of token selection, but it does not specifically penalize repeated topics. The question asks for a parameter that discourages the model from repeatedly discussing the same subject matter, which is the role of presence penalty.

Frequency penalty reduces repetition of the same tokens or phrases, but the question asks for introducing new topics and avoiding repeated subject matter, which is controlled by presence penalty (penalizing tokens that have appeared at all).

A developer wants the chatbot to generate more creative and diverse responses, even if that means occasional repetition. Increasing temperature would be correct to boost randomness and novelty.

A developer wants to generate creative story ideas where each sentence should be diverse in vocabulary and phrasing, but without penalizing repetition of themes. Increasing top_p would allow more varied word choices while still permitting the same topics to be revisited.

If the question were: 'A developer wants to reduce the likelihood of the model repeating the same word or phrase multiple times in a single response, which parameter should be increased?' then frequency penalty would be correct.

Candidates may confuse 'introducing new topics' with 'increasing randomness', assuming higher temperature leads to more varied content, but it doesn't directly address repetition of the same subject.

Candidates may confuse top_p with presence penalty because both influence output diversity, but top_p affects token selection probability distribution rather than directly penalizing topic repetition.

Candidates may confuse 'repeatedly discussing the same subject matter' with 'repeating the same words or phrases', not realizing that presence penalty targets topic-level novelty while frequency penalty targets token-level repetition.

Custom Vision requires training a custom model with labeled images, but the question specifies 'without needing to train a custom model.' Image Analysis provides pre-built captioning and classification capabilities out of the box.

Face Detection identifies human faces and their attributes (e.g., age, emotion) but does not generate descriptive captions or classify paintings as portrait/landscape.

OCR extracts text from images, not descriptive captions or painting classifications (portrait vs. landscape). The question requires generating captions and classifying art, which OCR cannot do.

Custom Vision would be correct if the museum needed to classify paintings into custom categories (e.g., by artist or period) that are not covered by Azure's pre-built models, and they are willing to train a model with their own labeled images.

A question asking which Azure service can detect faces in an image, return face rectangles, and optionally analyze attributes like smile or glasses, without needing custom training.

A question asking: 'Which Azure service extracts printed or handwritten text from images, such as reading labels on museum artifacts?' would make OCR correct.

Candidates may think that generating descriptive captions and classifying paintings requires custom training, not realizing that Azure's Image Analysis already includes pre-trained models for common tasks like captioning and art classification.

Candidates may think 'portrait' implies face detection is needed, but the task requires caption generation and classification, not just face detection.

Candidates may think OCR can 'read' the painting's content (e.g., title or artist signature) and infer its type, but OCR only extracts text, not visual features or scene understanding.

The question describes a bias in scholarship scoring based on major, which violates fairness, not reliability/safety. Reliability/safety concerns system failures or harm, not discriminatory outcomes.

The question describes bias in scholarship scores based on academic major, which is a fairness issue. Privacy and security relate to protecting personal data and preventing unauthorized access, not to biased outcomes.

The question describes a system producing biased outcomes against certain groups, which directly violates the Fairness principle. Transparency concerns explainability and disclosure, not the bias itself.

This option would be correct if the AI system produced inconsistent or incorrect scores due to data quality issues, such as missing or erroneous training data, leading to unreliable predictions that could cause financial harm to students.

This option would be correct if the AI system exposed students' personal information (e.g., grades, financial data) without consent or had a data breach, violating data protection laws.

A scenario where an AI system's decision-making process is hidden or not disclosed to users, such as a loan approval system that does not provide reasons for rejection, would make Transparency the correct answer.

Candidates may confuse 'unfair outcomes' with 'unreliable system' because bias can make a system seem unreliable, but reliability focuses on performance consistency and safety, not equity.

Candidates may confuse 'bias in data' with 'privacy violations' because both involve data handling, but bias is about fairness, not data protection.

Candidates may confuse the lack of fairness with a lack of transparency, thinking that if the system were more transparent, the bias would be exposed, but the core violation is still fairness.

Accuracy is not robust to class imbalance because a model that predicts the majority class for all instances can achieve 95% accuracy, failing to identify any positive cases.

Log loss measures the average log-likelihood of predictions, but it is not robust to class imbalance because it treats all errors equally, and the majority class dominates the loss, making it insensitive to improvements in minority class recall.

When the dataset is balanced (e.g., 50% positive, 50% negative) and the cost of false positives and false negatives is equal, accuracy is a straightforward and appropriate metric.

In a regression problem where the goal is to minimize the difference between predicted probabilities and actual outcomes, and the dataset is balanced, Log loss is the appropriate primary metric for probabilistic classification models.

Candidates often default to accuracy as the most familiar metric, not realizing that it is misleading for imbalanced datasets.

Candidates may confuse Log loss as a robust metric because it is commonly used in classification, but they overlook its sensitivity to class imbalance compared to AUC_weighted.

Decreasing Top_p reduces the set of tokens considered for sampling, which can make output more deterministic, but the question specifically asks for the parameter to decrease to reduce variability. Temperature is the primary parameter controlling randomness; decreasing it makes the model more deterministic. Top_p is an alternative sampling strategy, not the direct control for variability.

Frequency penalty reduces repetition of tokens based on their frequency in the generated text, not the logical correctness or determinism of the output. Decreasing it would not make the model's output more deterministic or reduce logical errors.

Presence penalty reduces the likelihood of repeating the same topic, not the randomness or variability of the output. Decreasing it would not make the model more deterministic.

A developer wants to ensure the model only considers the most likely tokens, avoiding rare or unusual words. Decreasing Top_p (e.g., from 0.9 to 0.1) would be correct when the goal is to limit the cumulative probability mass of token choices, making output more focused and less diverse.

A developer uses Azure OpenAI Service to generate creative stories, but notices the model repeats the same phrases too often. To reduce this repetition, which parameter should the developer increase? (Answer: Frequency penalty)

A developer uses Azure OpenAI Service to generate creative stories but notices the model keeps introducing new characters or topics. To discourage the model from talking about new topics and keep the story focused, which parameter should be increased?

Candidates may confuse Top_p with Temperature because both control randomness; they might think decreasing Top_p similarly reduces variability, not realizing Temperature is the primary parameter for that purpose.

Candidates may confuse frequency penalty with controlling randomness, assuming that penalizing frequent tokens makes output more predictable, but it actually targets repetition, not determinism.

Candidates may confuse 'penalty' parameters with controlling randomness, assuming that lowering a penalty reduces variability, but presence penalty specifically targets topic repetition, not output determinism.

Optical Character Recognition (OCR) extracts text from images, but it does not analyze visual content for violence or adult material. The question requires content moderation, which is not provided by OCR.

Object Detection identifies and locates objects in images (e.g., people, cars), but does not classify content as violent or adult. It lacks the content moderation labels needed for flagging inappropriate material.

Background Removal is used to isolate the foreground subject from the background, not to detect violent or adult content. It does not analyze image content for moderation purposes.

OCR would be correct if the question asked for extracting text from images, such as reading license plates from traffic camera images or digitizing printed documents for search indexing.

A question asking: 'Which Azure Computer Vision capability should be used to count the number of cars in an image and draw bounding boxes around them?' would make Object Detection the correct answer.

A question asking: 'Which Azure Computer Vision capability can automatically remove the background from product images for an e-commerce catalog?' would make Background Removal the correct answer.

Candidates may confuse OCR with general image analysis, thinking that detecting text could help identify inappropriate content like hate speech in images, but OCR only extracts text without understanding its context or detecting violent/adult imagery.

Candidates may think Object Detection can recognize violent or adult content because it detects objects like weapons or people, but it does not provide the safety classification labels required for content moderation.

Candidates might think Background Removal can filter out inappropriate backgrounds, but it lacks the content classification needed for moderation.

The question specifically asks about avoiding unfair discrimination based on age, gender, or ethnicity, which directly relates to the Fairness principle. Reliability and safety focuses on system dependability and risk mitigation, not on bias or discrimination.

The question focuses on preventing unfair discrimination based on age, gender, or ethnicity, which is directly addressed by the Fairness principle. Privacy and security relate to protecting data from unauthorized access or misuse, not to ensuring equitable outcomes across demographic groups.

Inclusiveness focuses on ensuring the system works for people of all abilities and backgrounds, but the question specifically asks about avoiding unfair discrimination based on age, gender, or ethnicity, which is directly addressed by the Fairness principle.

This option would be correct in a scenario where the bank's AI system must consistently perform correctly under varying conditions, such as handling unexpected input data or system failures, to ensure loan decisions are reliable and safe.

This option would be correct for a question like: 'A bank is developing an AI system to process loan applications and wants to ensure customer financial data is encrypted and access is restricted. Which principle applies?'

A question asks: 'Which Microsoft responsible AI principle ensures that AI systems are designed to be usable by people with a wide range of abilities, including those with disabilities?' In that context, Inclusiveness is the correct answer.

Candidates may confuse fairness with reliability, thinking that a reliable system inherently avoids bias, or they may overgeneralize the importance of reliability in all AI systems without recognizing the specific focus on discrimination in this question.

Candidates may confuse fairness with privacy, thinking that protecting demographic data prevents discrimination, but fairness requires active mitigation of bias, not just data protection.

Candidates may confuse 'inclusiveness' with 'fairness' because both involve treating people equitably, but inclusiveness is broader and includes accessibility, while fairness specifically targets bias and discrimination.

The question specifies 'without requiring any custom training,' but Custom Named Entity Recognition requires labeled training data to build a custom model, making it unsuitable for this zero-shot medical entity extraction task.

Key Phrase Extraction identifies general key terms (e.g., 'patient', 'treatment') but is not designed to extract medical-specific entities like disease names, medications, and dosages from free-text clinical notes.

Sentiment Analysis determines the emotional tone (positive, negative, neutral) of text, not medical entities like diseases or medications. The question specifically requires extracting medical entities from free-text patient notes, which is outside Sentiment Analysis's scope.

A hospital needs to extract unique entity types (e.g., internal department codes, local abbreviations) from patient notes that are not covered by pre-built models. The question would state that labeled training data is available and a custom solution is required.

A question asks: 'Which Azure AI Language feature extracts the most important points or topics from a document without any customization?' Key Phrase Extraction would be correct for summarizing main themes in general text.

A social media monitoring company wants to automatically classify customer feedback about a new product as positive, negative, or neutral. They need a pre-built feature that does not require custom training. Sentiment Analysis would be the correct answer.

Candidates may confuse 'custom' with 'specialized' and think that medical entities require a custom approach, overlooking that Text Analytics for Health is a pre-built, specialized model for healthcare.

Candidates may confuse 'extracting important phrases' with 'extracting medical entities,' assuming Key Phrase Extraction can handle medical terms without realizing it lacks domain-specific medical entity recognition.

Candidates may confuse 'analyzing text' with 'extracting entities' and think Sentiment Analysis can identify medical terms, or they may not be aware of the specialized Text Analytics for Health feature.

The actions described focus on assigning responsibility and reviewing adverse outcomes, which directly implement accountability, not fairness. Fairness addresses bias and equitable treatment, which is not the primary concern here.

The actions described—logging, reviewing adverse outcomes, and assigning responsibility—directly address accountability, not reliability and safety. Reliability and safety focus on system performance and risk mitigation, not on assigning human responsibility for outcomes.

The question describes logging, review of adverse outcomes, and assigning responsibility to a committee—these are accountability mechanisms, not transparency. Transparency focuses on providing explanations or understandable information about how the AI works, which is not mentioned here.

A question describing an AI system that produces different diagnostic accuracy for different skin tones, and the company implements measures to detect and mitigate such disparities, would make Fairness the correct answer.

A scenario where the AI system provides explanations for its predictions, such as a loan approval system that shows which factors influenced the decision, and the company publishes a plain-language description of the model's purpose and limitations.

Candidates may confuse accountability with fairness because both involve oversight and review, but fairness specifically targets bias and discrimination, not general responsibility for outcomes.

Candidates may confuse accountability with reliability and safety because both involve oversight and review processes. They might think that logging and reviewing adverse outcomes is about ensuring system reliability, but the key distinction is that accountability focuses on assigning human responsibility, not just system performance.

Candidates may confuse accountability with transparency because both involve oversight and documentation, but transparency specifically requires clear communication about the system's behavior and limitations, not just responsibility assignment.

Accountability refers to the principle that AI systems should be owned and overseen by people who can be held responsible for their outcomes. In this scenario, the primary concern is ensuring the chatbot serves diverse populations fairly across multiple languages, which directly relates to inclusiveness, not accountability.

The question focuses on serving diverse populations fairly across multiple languages, which directly relates to inclusiveness. Privacy and security, while important, are not the primary principle for ensuring fair service to diverse groups.

The question focuses on serving diverse populations fairly across multiple languages, which directly relates to inclusiveness, not transparency. Transparency is about ensuring users understand how AI systems work, not about fairness across demographics.

A company develops a chatbot that collects personal data (e.g., names, addresses) and wants to ensure this data is protected from unauthorized access. The question would ask: 'Which principle should they prioritize to safeguard customer information?'

A company deploys an AI system that makes decisions affecting customers (e.g., loan approvals). The exam asks which principle ensures customers can understand how decisions are made and challenge them. Transparency would be the correct answer.

Candidates may confuse accountability with inclusiveness because both involve ethical considerations. They might think that being accountable for serving diverse populations is the same as ensuring inclusiveness, but accountability is about ownership and responsibility, not about designing for diversity.

Candidates may confuse data protection with fairness, assuming that privacy is necessary for serving diverse populations, but the question explicitly asks about serving fairly, not about data handling.

Candidates may confuse transparency with fairness, thinking that being open about AI processes inherently ensures fair treatment of diverse groups, but transparency alone does not guarantee inclusiveness.

The question focuses on system performance under varying factory conditions (temperature, noise), which directly relates to reliability and safety, not fairness. Fairness addresses bias against groups, not environmental robustness.

The question focuses on performance under varying factory floor conditions (temperature, noise), which directly relates to system reliability and safety, not to protecting data or preventing unauthorized access.

Inclusiveness focuses on designing AI systems that are accessible and fair to all users, regardless of background or ability. It does not address performance under varying environmental conditions like temperature and noise.

An AI system for loan approvals must ensure it does not discriminate based on race or gender. The question would ask which principle ensures unbiased outcomes across demographic groups.

A question that asks which principle addresses protecting sensitive machine data or preventing unauthorized access to the AI system would make Privacy & Security the correct answer.

A question about an AI system that must work effectively for users with diverse abilities, languages, or cultural backgrounds, such as a voice assistant that must understand different accents or a hiring tool that must avoid bias against underrepresented groups.

Candidates may confuse 'fairness' with general system robustness, thinking that reliable performance under all conditions is a form of fairness to all users.

Candidates may confuse operational robustness with data protection, or think that varying conditions introduce security risks, leading them to select Privacy & Security instead of Reliability & Safety.

Candidates may confuse inclusiveness with robustness, thinking that including diverse conditions means the system should work under all conditions, but inclusiveness specifically relates to human diversity, not environmental variability.

MAE does not heavily penalize large errors because it treats all errors linearly, whereas the question requires a metric that penalizes large errors more, which is a key property of MSE.

Classification Accuracy is used for classification tasks, not regression. The question asks about predicting a continuous number (rentals per hour), so accuracy is inappropriate.

R-squared measures the proportion of variance explained by the model, not the penalty for large errors. It does not heavily penalize occasional large errors like MSE does.

A bike-sharing company wants a metric that is easy to interpret in the original unit (number of rentals) and is robust to outliers. They would use MAE because it gives the average absolute error without disproportionately penalizing occasional large errors.

For a question like 'Which metric should be used to evaluate a model that predicts whether a bike will be rented (yes/no) in the next hour?', Classification Accuracy would be correct.

A question asking which metric indicates how well the independent variables explain the variability of the dependent variable in a regression model, e.g., 'Which metric shows the proportion of variance in the target explained by the features?'

Candidates may think MAE is always the best choice for regression because it is intuitive and less sensitive to outliers, but they overlook the specific requirement to heavily penalize large errors.

Candidates may confuse regression with classification or think 'accuracy' applies broadly to any prediction task.

Candidates may confuse R-squared as a metric for error magnitude because it is commonly used in regression evaluation, but it does not directly penalize large errors.

Increasing temperature makes output more random, which could introduce more repetition or variation, not reduce it. Repetition is controlled by frequency penalty, which penalizes tokens that have already appeared.

Increasing max tokens allows longer output but does not reduce repetition; it may even increase repetition by giving the model more space to generate redundant content.

Top P controls the cumulative probability of token selection, affecting diversity but not directly penalizing repetition. Increasing Top P would not specifically reduce repetitive patterns like repeated function definitions or loops.

A question asking which parameter to increase to make the model generate more creative or diverse responses, e.g., 'Which parameter should be increased to encourage the model to produce less predictable code?'

A question asks: 'Which parameter should be increased to allow the model to generate a longer response, such as a complete essay or a full code file?'

In a scenario where the question asks which parameter to adjust to make the generated text more focused and less random, such as 'Which parameter should be decreased to produce more deterministic and less diverse outputs?', then Top P would be correct if decreased.

Candidates may confuse temperature with frequency penalty, thinking that higher randomness (temperature) will break repetitive patterns, but it actually increases the chance of random repetitions.

Candidates may think that limiting output length causes repetition, so increasing max tokens would give the model room to avoid repeating itself, but repetition is controlled by penalty parameters, not length.

Candidates may confuse Top P with frequency penalty because both influence output diversity, but Top P affects the nucleus sampling threshold rather than directly penalizing repeated tokens.